rsc_fpops_est / GPU

[rebi...@web.de]

Hi,

 

we have a discussion about this value in WU input_template for GPUs. The estimated runtime in BOINC is too high or too low

 

rsc_fpops_est / FLOPs

 

For CPU all is correct with this formula but for GPU it is totally wrong.

 

@David: DCF could be solution but only for one subproject (GPU). We are using fixed credits with the old credit system, quorum 1 without DCF.

 

-Reb

 

 

[David P. Anderson]

Where in the code are you talking about (client or server)?

In calculations of that sort, FLOPS is different for CPU and GPU app versions.

If you're using plan classes for your GPU apps,

the FLOPS should be estimated correctly.

--
You received this message because you are subscribed to the Google Groups "boinc_projects" group.
To unsubscribe from this group and stop receiving emails from it, send an email to boinc_project...@ssl.berkeley.edu.
To view this discussion visit https://groups.google.com/a/ssl.berkeley.edu/d/msgid/boinc_projects/0C09B4DD29CB417283F05D1D15A09403%40Testserver.

[rebi...@web.de]

Hello David,

 

both, client and server

 

for server started with entry in input_template

<rsc_fpops_est>8e11</rsc_fpops_est>

 

We have a plan_class for AMD, Intel ARC and Nvidia apps for linux and windows.

 

But how can recognise the client to use the correct estimated runtime? The plan_class has a <gpu_peak_flops_scale> value. Generally what do we need to change or add?

 

-Reb